Skip to content

[Docs] Note pooled aggregation semantics for mean_acceptance_length #52169

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
arrdel wants to merge 1 commit into
base: main
Choose a base branch
from
+12 −1
Open

[Docs] Note pooled aggregation semantics for mean_acceptance_length #52169

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 12 additions & 1 deletion vllm/v1/spec_decode/metrics.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -110,7 +110,11 @@ def log(self, log_fn=logger.info):
else float("nan")
)

# Conventionally, mean acceptance length includes the bonus token
# Conventionally, mean acceptance length includes the bonus token.
# This is a *pooled* mean (weighted by drafts): it is equivalent to
# total_generated / total_verify_calls, not the unweighted mean of
# per-request means that some external benchmarks report (see #42508
# for a reproduction against SpecForge on variable-length workloads).
mean_acceptance_length = 1 + (num_accepted_tokens / num_drafts)

pos_matrix = np.array(self.accepted_tokens_per_pos_lists)
Expand Down Expand Up @@ -189,6 +193,13 @@ class SpecDecodingProm:
rate(vllm:spec_decode_num_accepted_tokens_total[$interval]) /
rate(vllm:spec_decode_num_drafts[$interval]))

Both the log line and this PromQL formula compute a *pooled* mean,
weighted by draft count. External benchmarks that instead report an
unweighted mean of per-request means (e.g., SpecForge's
``average_acceptance_length``, which is ``statistics.fmean`` over
per-request lengths) are not directly comparable on variable-length
workloads. See #42508.

A per-position acceptance rate vector can be computed using

vllm:spec_decode_num_accepted_tokens_per_pos[$interval] /
Expand Down
Loading